Generative Grammar 295 12 Generative Grammar THOMAS WASOW 1 Introduction 1.1 “Grammar” To most people, the word “grammar” suggests a normative enterprise, dictating what constitutes correct language use. For example, many educated English speakers would identify at least five supposed grammatical “errors” in the following sentence: (1) Hopefully, we will be able to easily figure out who to talk to. Yet native speakers of American English also certainly recognize that (1) would be an entirely acceptable and natural sounding sentence in ordinary discourse. Indeed, the supposedly “correct” alternative (2) would be an awkward and affected way of expressing the thought. (2) I hope that we shall be able easily to figure out to whom to talk. Modern linguistics has little use for this prescriptive conception of grammar. Linguists are more interested in the knowledge of English that allows native speakers to judge (1) as fully acceptable and (2) as somewhat less natural. The prescriptions of traditional grammar are attempts to impose the speech pat- terns of one region, class, ethnicity, or generation on speakers belonging to other groups. They may be of interest to sociologists, historians, and political scient- ists, but they tell us very little about the nature of language. Language is a natural phenomenon, constituting an essential component of every human society. Linguistics is concerned with studying languages and language in general, much as biology studies living species and life in gen- eral. From this scientific perspective, the norms of prescriptive grammar are to 296 Thomas Wasow linguistics as the American Kennel Club’s breed standards are to biology: arbitrary evaluative standards of no relevance to objective description. Linguists use the term “grammar,” then, to refer to structural properties of language that have evolved naturally and that native speakers of the language have mastered without explicit instruction. These are largely properties of languages that are not even mentioned in traditional grammars, though some are addressed in foreign language instruction. They include facts about word order, for example, that we, will, and be in (1) must appear in that order, or else the sentence becomes unacceptable. They also include facts about the proper forms of words in particular contexts, for example, that replacing figure in (1) with figured, figures, or figuring makes the sentence unacceptable. Put in more technical jargon, “grammar” is taken by linguists to encompass syntax and morphosyntax. The term may also be construed more broadly to include principles relating linguistic forms to the meanings they express (semantics) and / or the sound patterns of languages (phonology). 1.2 “Generative” The term “generative” is associated with the tradition of grammatical research initiated and inspired by the work of Noam Chomsky. This term is some- times construed very narrowly to refer only to work directly derivative from Chomsky’s. Here it will be used more broadly to refer to work generally within the Chomskyan tradition, irrespective of whether its formalism and terminology come directly from Chomsky. Among Chomsky’s most important insights is the observation (noted inde- pendently over a century earlier by the great German linguist Wilhelm von Humboldt) that there are infinitely many well-formed sentences in any natural language. This follows immediately from the fact that any limit one might impose on the length of sentences would be arbitrary: any supposedly longest English sentence S would be two words shorter than “I said S,” which is surely well-formed if S is. On the other hand, a grammar, conceived of as a descrip- tion of a language, should be finite. How can we give a finite description of something infinite? Inspired by earlier work in mathematical logic and the foundations of computer science, Chomsky answered this question by proposing that we think of grammars as devices that put pieces of sentences together according to precise rules, thereby “generating” well-formed sentences. If some of the grammar rules can apply to their own outputs (in technical jargon, if some rules are “recursive”), then it is possible for finite grammars to generate infinite languages. To illustrate this, consider the following very simple (nonlinguistic) example. The ordinary Arabic numeral system used to represent numbers has infinitely many well-formed expressions (one for each number) constructed out of ten symbols, namely, the digits “0” through “9.” We can write a simple grammar for the numerals denoting positive integers with the following rules: Generative Grammar 297 • Each of the digits 1, 2, 3, 4, 5, 6, 7, 8, and 9 is a numeral. • If N is any numeral, then N0 is a numeral. • If N is any numeral, then NN is a numeral. One of many possible formalizations of this would be the following: N→1N→5N→9 N→2N→6N→N0 N→3N→7N→NN N→4N→8 Here N is the category of well-formed numerals, and the arrow can be inter- preted to mean “may consist of.” This little grammar generates the infinite “language” of numerals denoting positive integers, because it contains rules that are recursive (namely, the last two). 2 Tenets of Generative Grammar Although the term “generative” orginally characterized a conception of gram- mars as such recursive rule systems, the term is now used somewhat more generally. In particular, what distinguishes work in generative grammar is the goal of describing languages systematically, as opposed to the more anecdotal approach of traditional grammars. While it is impossible to give a precise definition of generative grammar, there are several tenets shared by the vast majority of generative grammarians. These are summarized in the following subsections. 2.1 Grammars should be descriptive, not prescriptive As discussed above, this proposition is generally accepted by modern linguists. Although it is not unique to generative grammarians, it is common to them. 2.2 Grammars should characterize competence, not performance Despite its anti-prescriptivism, generative grammar is not an attempt to describe all or only the actual utterances of native speakers. This is implicit in the claim that languages are infinite: it would have been safe to assume that no sentence over one million words long will ever be uttered. But this upper bound exists because of limits on human memory and patience, not because of any linguistically interesting facts. Moreover, because of speech errors of various kinds, people frequently produce utterances that are not well-formed 298 Thomas Wasow sentences, even by the judgments of the speakers. To distinguish between the idealized infinite languages that generative grammarians seek to describe and the far messier output of actual speakers, Chomsky introduced the terminology “competence” vs. “performance.” One common property of generative grammar in all its varieties is the focus on characterizing linguistic competence. Many generative grammarians would also like to develop models of linguistic performance, but most believe that a competence theory will be a necessary component of such a model. Put slightly differently, it is widely accepted that explaining how a language is actually used will require understanding speakers’ knowledge of that language. 2.3 Grammars should be fully explicit Traditional grammars presuppose some knowledge of the language under de- scription and tend to focus on aspects of the language that are variable or have changed. Generative grammars are supposed to be precise rule systems that characterize the whole language, without relying on any prior knowledge of the language on the part of the reader. Many generative grammarians identify explicitness with formalization. Hence, the generative literature abounds with formalisms (though it is not always made clear how the formalisms are to be interpreted). Early work in generative grammar approached this goal of ex- plicitness and formalization far more consistently than most recent work. 2.4 Linguistic analyses should be maximally general If two grammars cover the same range of data, but one requires two distinct rules where the second has only one, generative grammarians take this as evidence for the superiority of the second grammar. A famous example of this mode of reasoning is due to Postal (1964). He noted that what are called “tag questions” in English require a kind of matching between the tag and the initial portions of the main clause, as illustrated in (3). Following standard practice, asterisks are used to mark unacceptable strings. I haven’t *you (3) a. I have won, *won’t ? *we *aren’t *they won’t you b. You will win, *haven’t *we ? *aren’t *they Postal also observed that imperative sentences take only a restricted range of tags, though there is nothing overtly present in the initial portions of imper- ative sentences that the tags match. Generative Grammar 299 you won’t *I (4) Close the door, *haven’t ! *we *aren’t *they If we analyze imperative sentences as having an initial you will at some level of analysis, he reasoned, we could use a simple rule to generate tag questions on both declarative and imperative sentences. Such an analysis is said to “capture a generalization” – in this case, the generalization that tags on imperatives and declaratives are fundamentally alike. The desire to capture generalizations plays a very important role in the argumentation of generative grammar. 2.5 The theory of grammar should make universal claims To the extent possible, facts about individual languages should be derived from general principles that apply to all languages. Information stipulated in the grammars of particular languages should be kept to a minimum. This is motivated in part simply by standard scientific methodological considerations: more general hypotheses are both more parsimonious and more interesting than less general ones. But it is also motivated in part by psychological concerns – specifically, by Chomsky’s “argument from the poverty of the stimulus,” which will be discussed in the next subsection.